hello,there is some wrong with the compute_adjustment

[limengyang1992]

There is no problem with 10 classification, when 100 classification, the order is wrong
I modified it：

label_freq = dict(sorted(label_freq.items()))
label_freq = sorted(label_freq.items(), key=lambda e: e[1], reverse=True)
label_freq_array = np.array([x[1] for x in label_freq])
label_freq_array = label_freq_array / label_freq_array.sum()

[Chumsy0725]

Hey! @limengyang1992
It can be rectified by casting the keys as integers.

    for j in target:
       key = int(j.item())
       label_freq[key] = label_freq.get(key, 0) + 1

It is convenient to have adjustments sorted according to the class number than the class frequency.
Thanks!

[sadeepj]

Hi @Chumsy0725,

I didn't have a deeper look - but it looks like there's still a problem here. As seen from the screenshot, the tensor returned by the above compute_adjustment() method has the order ['0', '1', '10', '11', ...] instead of ['0', '1', '2', '3', ...]. Wouldn't this cause a problem for CIFAR-100 when this output tensor is used to correct the logits, which are always ordered ['0', '1', '2', '3', ...] ?

[Chumsy0725]

Hi Dr. @sadeepj ,

The issue of keys being out of order is corrected by casting the key as integers. The resulting label_freq dictionary is attached bellow.

also the resulting probability values (attached bellow) in label_freq_array after normalizing are the same as the values listed in google-research/logit_adjustment and the values are in the same order.

[sadeepj]

Thanks @Chumsy0725! I didn't realize that the issue is already fixed by 2e5bd39.

@limengyang1992, does this solve the issue you were facing?